All-In — Gerstner: no AI bubble; semis = ~70% of Nasdaq return

Opening

No classic multiple bubble — the AI trade is earnings- and offtake-constrained, with semis already ~70% of Nasdaq's return and lab monthly revenue as the takeoff switch. Brad Gerstner (Altimeter) delivers an ~18m All-In Summit speed-round market/tech check after Trump Accounts / CAC-scan openers. Ground covered: YTD scoreboard vs gold/BTC; Mag7/SOX multiples vs 2000; hyperscaler capex ≈ semi FCF; Anthropic revenue "fuse"; offtake path to fund Mag5 build; Dylan Patel 43 GW vs Gerstner's ~25 GW; regulation / nuclear precedent; rates as gravity. Fidelity is asr — proper nouns and lab revenue run-rates marked. Watch

Key takes

Gerstner frames 2026 as an earnings-driven expansion with multiple contraction, not a 2000-style bubble. Markets up ~15% YTD and ~39% since Jan of prior year despite tariffs/geopolitics/AI-regulation worry; gold flat, Bitcoin ~−10%; NVDA revenue, hyperscaler capex, and OpenAI/Anthropic valuations each ~2×; SpaceX ~2.5×. Nasdaq/S&P/SOX/NVDA all "well below" average multiples; Mag7 roughly in-line; consumer discretionary, software, and financials "barely moved." [book]

Semiconductors are ~70% of the Nasdaq's return — concentration as both feature and risk. "Makers of the tokens" capture the money; buyers ride along because infrastructure is tight. Dell cited as public company with VC-like returns (~5× / ~9× in ~18 months). Chart claim: hyperscaler capex is almost dollar-for-dollar free cash flow to the semiconductor infrastructure complex. [book]

The load-bearing offtake question: who pays the rent on Mag5 build? Gerstner revisits his Oct prior-year Sam Altman question ($1T capex commit vs ~$13bn GAAP revenue) — then points to Anthropic's Dec→Feb→Mar monthly revenue step-up as the "exclamation point" that lit the April/May rip, with June/July consolidation after ARR revised (~65 vs street ~75) plus open-source catch-up worry. Collective top-lab run-rate framed ~$100bn (asr — names Anthropic, OpenAI, and "SpaceX"; treat third name as uncertain); needs ~+$80bn to ~$180bn by year-end to "keep the AI trade intact." Mag5 builds to rent; offtake must climb ~200 → 450 → 800/1,000bn or the capex path does not clear. [book]

Supply: SemiAnalysis 43 GW next year is too aggressive — Gerstner's hunch ~25 GW, half to Anthropic + OpenAI. This year ~19 GW added with ~7 to the two leading labs; Dylan Patel forecast +43 GW / ~14 to leaders next year — equal to cumulative US compute today (<~40 GW). Permitting, local opposition, interconnection, labor, sold-out power equipment bind. Revenue math: Anthropic reportedly doing ~$100–110bn this year on ~1.5 GW; +4–5 GW "certainly enough" for another ~$100bn — so GW miss ≠ automatic revenue miss for next year. By 2028, >half of US compute controlled by two labs (citing Sacks). [book]

TAM / tokens / margin expansion are the bull stack; regulation, power, and rates are the three risks. Knowledge-work TAM only needs 4% ($1.2T) to pay for capex; ~47 quadrillion tokens; Codex users ~40× in eight months; median enterprise AI spend ~17× in 18 months. Nasdaq 2015–25 EPS ~10% (~6% rev + ~38 bps margin/year) — can AI lift margin expansion toward ~100 bps (Uber/Snowflake "grow without headcount" anecdotes). Risks: (1) activist-driven over-regulation (nuclear: ~67 fission reactors shut — "unilateral disarmament" vs China); (2) atoms/energy hard; (3) rate hikes / higher DC hurdle rates; 10y toward ~5.5% as equity gravity. [book]

Portfolio stance for 2026: "medium," mentally flexible — lab monthly revenues are the single most important datapoint. 2023–25 only required being long the AI supercycle; 2026 is "facts and circumstances," already priced, "don't YOLO." If lab monthly revenues print closer to ~$8bn than ~$4bn and oil/rates cooperate → add chips; else reserve right to go smaller. Anthropic IPO halt/postponement on regulation is named as a major downside path he does not base-case. [book]

Key math

Nasdaq ~+15% YTD; ~+39% since Jan prior year; earnings ~+26%; multiples down (asr) — earnings tape, not multiple bubble. [book]

NVDA ~14× next-year fully taxed GAAP EPS; semis ~70% of Nasdaq return (asr) — concentration + valuation discipline claim. [book]

Hyperscaler capex ≈ dollar-for-dollar semi-infra FCF; Dell ~5× / ~9× in ~18 months (asr) — who harvests the build. [book]

Anthropic monthly path Dec ~$2bn → Feb ~$4bn → Mar ~$11bn; ARR revised ~$65bn vs ~$75bn expected (asr — Gerstner/rumor framing) — offtake "fuse" narrative. [book]

Top-lab collective RR ~$100bn → need $180bn by YE (+$80bn); offtake path ~$200bn → $450bn → $800bn/$1T (asr) — rent math for Mag5 capex. [book]

Compute: ~19 GW added 2026 (~7 to two labs); Dylan +43 GW next (~14 to labs); Gerstner ~25 GW (half to Anthropic+OpenAI); US cumulative compute <~40 GW (asr) — power/stand-up bind. [book]

Anthropic ~$100–110bn RR this year on ~1.5 GW; +4–5 GW → another ~$100bn revenue capacity (asr) — GW shortfall ≠ revenue shortfall next year. [book]

~4% of knowledge-work TAM ≈ $1.2T pays for capex; ~47 quadrillion tokens; Codex users ~40× / 8mo; enterprise median spend ~17× / 18mo; margin expansion ~38 bps → ?100 bps (asr) — demand / productivity stack. [book]

Quotes

"This is no bubble like it was in 2000." — Brad Gerstner [asr] [book]

"Semiconductors are 70% of the Nasdaq's… return." — Brad Gerstner [asr] [book]

"Microsoft's not paying for it. They're building it to rent it." — Brad Gerstner [asr] [book]

"I would suggest Dylan's forecast to 43 gigawatt next year is too aggressive… closer to 25 gigawatt." — Brad Gerstner [asr] [book]

"The single most important data point in the market today… is Anthropic's monthly revenue [and] OpenAI's monthly revenue going to be 4 billion or 8 billion?" — Brad Gerstner [asr] [book]

"That is not where we are in 2026. Everybody knows about AI. It's all priced now. It's about facts and circumstances… Don't YOLO." — Brad Gerstner [asr] [book]

Variant perception

Priced in — AI as the supercycle; Mag7/hyperscaler capex durability; NVDA as core expression; "is it a bubble" as the standing debate. Post-Fed week already trading concentration and rates.

What's new — explicit offtake-vs-rent grammar plus a falsifiable GW haircut (43→~25) that still clears next-year lab revenue if Anthropic/OpenAI get ~half. Monthly lab RR as the takeoff/abort switch, and semis-as-70%-of-Nasdaq-return as a concentration risk statement from a known AI bull, are the desk-usable bits. Nuclear-regulation analogy is the political-risk framing.

Bear case — if monthly lab RR stalls near ~$4bn, open-source compresses pricing, or 10y pushes toward ~5.5%, his own fan chart goes lower even without a "bubble pop." 70% Nasdaq contribution from semis is fragile to any offtake miss. Third "lab" in the $100bn collective RR line is ASR-noisy. Summit charts are Altimeter marketing collateral.

Discount — Gerstner manages Altimeter and is a structural AI/semi bull talking his book on stage at All-In Summit. Partner reads (IREN, Keel, etc.) sit under the bumper. Treat Anthropic/OpenAI monthly and ARR figures as rumor-composite unless corroborated. "SpaceX" in the top-lab RR list may be ASR error.

Positioning

AI capex durability — STRENGTHENS (conditionally). Offtake must fund Mag5 rent, but Gerstner's path still assumes multi-hundred-billion lab RR and continued build; GW haircut is a stand-up constraint, not a demand cancel. Conditional on monthly lab revenues printing toward the ~$8bn takeoff case.

HBM supply binds — NEUTRAL. Episode is offtake/power/rates, not memory hierarchy — though token/GW math sits upstream of accelerator shipments.

Inference margin inversion — NEUTRAL / watched. Open-source catch-up named as a June/July consolidation factor; no lab gross-margin path across price cuts disclosed here.